/* Subroutine */ int igraphdnaitr_(integer *ido, char *bmat, integer *n, integer *k, integer *np, integer *nb, doublereal *resid, doublereal *rnorm, doublereal *v, integer *ldv, doublereal *h__, integer *ldh, integer * ipntr, doublereal *workd, integer *info) { /* Initialized data */ static logical first = TRUE_; /* System generated locals */ integer h_dim1, h_offset, v_dim1, v_offset, i__1, i__2; doublereal d__1, d__2; /* Builtin functions */ double sqrt(doublereal); /* Local variables */ static integer i__, j; static real t0, t1, t2, t3, t4, t5; static integer jj, ipj, irj, ivj; static doublereal ulp, tst1; extern doublereal igraphddot_(integer *, doublereal *, integer *, doublereal *, integer *); static integer ierr, iter; static doublereal unfl, ovfl; static integer itry; extern doublereal igraphdnrm2_(integer *, doublereal *, integer *); static doublereal temp1; static logical orth1, orth2, step3, step4; static doublereal betaj; extern /* Subroutine */ int igraphdscal_(integer *, doublereal *, doublereal *, integer *), igraphdgemv_(char *, integer *, integer *, doublereal *, doublereal *, integer *, doublereal *, integer *, doublereal *, doublereal *, integer *); static integer infol; extern /* Subroutine */ int igraphdcopy_(integer *, doublereal *, integer *, doublereal *, integer *), igraphdaxpy_(integer *, doublereal *, doublereal *, integer *, doublereal *, integer *), igraphdmout_(integer *, integer *, integer *, doublereal *, integer *, integer *, char *); static doublereal xtemp[2]; extern /* Subroutine */ int igraphdvout_(integer *, integer *, doublereal *, integer *, char *); static doublereal wnorm; extern /* Subroutine */ int igraphivout_(integer *, integer *, integer *, integer *, char *), igraphdgetv0_(integer *, char *, integer *, logical *, integer *, integer *, doublereal *, integer *, doublereal *, doublereal *, integer *, doublereal *, integer * ), igraphdlabad_(doublereal *, doublereal *); static doublereal rnorm1; extern doublereal igraphdlamch_(char *); extern /* Subroutine */ int igraphdlascl_(char *, integer *, integer *, doublereal *, doublereal *, integer *, integer *, doublereal *, integer *, integer *); extern doublereal igraphdlanhs_(char *, integer *, doublereal *, integer *, doublereal *); extern /* Subroutine */ int igraphsecond_(real *); static logical rstart; static integer msglvl; static doublereal smlnum; /* %----------------------------------------------------% */ /* | Include files for debugging and timing information | */ /* %----------------------------------------------------% */ /* \SCCS Information: @(#) */ /* FILE: debug.h SID: 2.3 DATE OF SID: 11/16/95 RELEASE: 2 */ /* %---------------------------------% */ /* | See debug.doc for documentation | */ /* %---------------------------------% */ /* %------------------% */ /* | Scalar Arguments | */ /* %------------------% */ /* %--------------------------------% */ /* | See stat.doc for documentation | */ /* %--------------------------------% */ /* \SCCS Information: @(#) */ /* FILE: stat.h SID: 2.2 DATE OF SID: 11/16/95 RELEASE: 2 */ /* %-----------------% */ /* | Array Arguments | */ /* %-----------------% */ /* %------------% */ /* | Parameters | */ /* %------------% */ /* %---------------% */ /* | Local Scalars | */ /* %---------------% */ /* %-----------------------% */ /* | Local Array Arguments | */ /* %-----------------------% */ /* %----------------------% */ /* | External Subroutines | */ /* %----------------------% */ /* %--------------------% */ /* | External Functions | */ /* %--------------------% */ /* %---------------------% */ /* | Intrinsic Functions | */ /* %---------------------% */ /* %-----------------% */ /* | Data statements | */ /* %-----------------% */ /* Parameter adjustments */ --workd; --resid; v_dim1 = *ldv; v_offset = 1 + v_dim1; v -= v_offset; h_dim1 = *ldh; h_offset = 1 + h_dim1; h__ -= h_offset; --ipntr; /* Function Body */ /* %-----------------------% */ /* | Executable Statements | */ /* %-----------------------% */ if (first) { /* %-----------------------------------------% */ /* | Set machine-dependent constants for the | */ /* | the splitting and deflation criterion. | */ /* | If norm(H) <= sqrt(OVFL), | */ /* | overflow should not occur. | */ /* | REFERENCE: LAPACK subroutine dlahqr | */ /* %-----------------------------------------% */ unfl = igraphdlamch_("safe minimum"); ovfl = 1. / unfl; igraphdlabad_(&unfl, &ovfl); ulp = igraphdlamch_("precision"); smlnum = unfl * (*n / ulp); first = FALSE_; } if (*ido == 0) { /* %-------------------------------% */ /* | Initialize timing statistics | */ /* | & message level for debugging | */ /* %-------------------------------% */ igraphsecond_(&t0); msglvl = debug_1.mnaitr; /* %------------------------------% */ /* | Initial call to this routine | */ /* %------------------------------% */ *info = 0; step3 = FALSE_; step4 = FALSE_; rstart = FALSE_; orth1 = FALSE_; orth2 = FALSE_; j = *k + 1; ipj = 1; irj = ipj + *n; ivj = irj + *n; } /* %-------------------------------------------------% */ /* | When in reverse communication mode one of: | */ /* | STEP3, STEP4, ORTH1, ORTH2, RSTART | */ /* | will be .true. when .... | */ /* | STEP3: return from computing OP*v_{j}. | */ /* | STEP4: return from computing B-norm of OP*v_{j} | */ /* | ORTH1: return from computing B-norm of r_{j+1} | */ /* | ORTH2: return from computing B-norm of | */ /* | correction to the residual vector. | */ /* | RSTART: return from OP computations needed by | */ /* | dgetv0. | */ /* %-------------------------------------------------% */ if (step3) { goto L50; } if (step4) { goto L60; } if (orth1) { goto L70; } if (orth2) { goto L90; } if (rstart) { goto L30; } /* %-----------------------------% */ /* | Else this is the first step | */ /* %-----------------------------% */ /* %--------------------------------------------------------------% */ /* | | */ /* | A R N O L D I I T E R A T I O N L O O P | */ /* | | */ /* | Note: B*r_{j-1} is already in WORKD(1:N)=WORKD(IPJ:IPJ+N-1) | */ /* %--------------------------------------------------------------% */ L1000: if (msglvl > 1) { igraphivout_(&debug_1.logfil, &c__1, &j, &debug_1.ndigit, "_naitr: generat" "ing Arnoldi vector number"); igraphdvout_(&debug_1.logfil, &c__1, rnorm, &debug_1.ndigit, "_naitr: B-no" "rm of the current residual is"); } /* %---------------------------------------------------% */ /* | STEP 1: Check if the B norm of j-th residual | */ /* | vector is zero. Equivalent to determing whether | */ /* | an exact j-step Arnoldi factorization is present. | */ /* %---------------------------------------------------% */ betaj = *rnorm; if (*rnorm > 0.) { goto L40; } /* %---------------------------------------------------% */ /* | Invariant subspace found, generate a new starting | */ /* | vector which is orthogonal to the current Arnoldi | */ /* | basis and continue the iteration. | */ /* %---------------------------------------------------% */ if (msglvl > 0) { igraphivout_(&debug_1.logfil, &c__1, &j, &debug_1.ndigit, "_naitr: ****** " "RESTART AT STEP ******"); } /* %---------------------------------------------% */ /* | ITRY is the loop variable that controls the | */ /* | maximum amount of times that a restart is | */ /* | attempted. NRSTRT is used by stat.h | */ /* %---------------------------------------------% */ betaj = 0.; ++timing_1.nrstrt; itry = 1; L20: rstart = TRUE_; *ido = 0; L30: /* %--------------------------------------% */ /* | If in reverse communication mode and | */ /* | RSTART = .true. flow returns here. | */ /* %--------------------------------------% */ igraphdgetv0_(ido, bmat, &itry, &c_false, n, &j, &v[v_offset], ldv, &resid[1], rnorm, &ipntr[1], &workd[1], &ierr); if (*ido != 99) { goto L9000; } if (ierr < 0) { ++itry; if (itry <= 3) { goto L20; } /* %------------------------------------------------% */ /* | Give up after several restart attempts. | */ /* | Set INFO to the size of the invariant subspace | */ /* | which spans OP and exit. | */ /* %------------------------------------------------% */ *info = j - 1; igraphsecond_(&t1); timing_1.tnaitr += t1 - t0; *ido = 99; goto L9000; } L40: /* %---------------------------------------------------------% */ /* | STEP 2: v_{j} = r_{j-1}/rnorm and p_{j} = p_{j}/rnorm | */ /* | Note that p_{j} = B*r_{j-1}. In order to avoid overflow | */ /* | when reciprocating a small RNORM, test against lower | */ /* | machine bound. | */ /* %---------------------------------------------------------% */ igraphdcopy_(n, &resid[1], &c__1, &v[j * v_dim1 + 1], &c__1); if (*rnorm >= unfl) { temp1 = 1. / *rnorm; igraphdscal_(n, &temp1, &v[j * v_dim1 + 1], &c__1); igraphdscal_(n, &temp1, &workd[ipj], &c__1); } else { /* %-----------------------------------------% */ /* | To scale both v_{j} and p_{j} carefully | */ /* | use LAPACK routine SLASCL | */ /* %-----------------------------------------% */ igraphdlascl_("General", &i__, &i__, rnorm, &c_b25, n, &c__1, &v[j * v_dim1 + 1], n, &infol); igraphdlascl_("General", &i__, &i__, rnorm, &c_b25, n, &c__1, &workd[ipj], n, &infol); } /* %------------------------------------------------------% */ /* | STEP 3: r_{j} = OP*v_{j}; Note that p_{j} = B*v_{j} | */ /* | Note that this is not quite yet r_{j}. See STEP 4 | */ /* %------------------------------------------------------% */ step3 = TRUE_; ++timing_1.nopx; igraphsecond_(&t2); igraphdcopy_(n, &v[j * v_dim1 + 1], &c__1, &workd[ivj], &c__1); ipntr[1] = ivj; ipntr[2] = irj; ipntr[3] = ipj; *ido = 1; /* %-----------------------------------% */ /* | Exit in order to compute OP*v_{j} | */ /* %-----------------------------------% */ goto L9000; L50: /* %----------------------------------% */ /* | Back from reverse communication; | */ /* | WORKD(IRJ:IRJ+N-1) := OP*v_{j} | */ /* | if step3 = .true. | */ /* %----------------------------------% */ igraphsecond_(&t3); timing_1.tmvopx += t3 - t2; step3 = FALSE_; /* %------------------------------------------% */ /* | Put another copy of OP*v_{j} into RESID. | */ /* %------------------------------------------% */ igraphdcopy_(n, &workd[irj], &c__1, &resid[1], &c__1); /* %---------------------------------------% */ /* | STEP 4: Finish extending the Arnoldi | */ /* | factorization to length j. | */ /* %---------------------------------------% */ igraphsecond_(&t2); if (*(unsigned char *)bmat == 'G') { ++timing_1.nbx; step4 = TRUE_; ipntr[1] = irj; ipntr[2] = ipj; *ido = 2; /* %-------------------------------------% */ /* | Exit in order to compute B*OP*v_{j} | */ /* %-------------------------------------% */ goto L9000; } else if (*(unsigned char *)bmat == 'I') { igraphdcopy_(n, &resid[1], &c__1, &workd[ipj], &c__1); } L60: /* %----------------------------------% */ /* | Back from reverse communication; | */ /* | WORKD(IPJ:IPJ+N-1) := B*OP*v_{j} | */ /* | if step4 = .true. | */ /* %----------------------------------% */ if (*(unsigned char *)bmat == 'G') { igraphsecond_(&t3); timing_1.tmvbx += t3 - t2; } step4 = FALSE_; /* %-------------------------------------% */ /* | The following is needed for STEP 5. | */ /* | Compute the B-norm of OP*v_{j}. | */ /* %-------------------------------------% */ if (*(unsigned char *)bmat == 'G') { wnorm = igraphddot_(n, &resid[1], &c__1, &workd[ipj], &c__1); wnorm = sqrt((abs(wnorm))); } else if (*(unsigned char *)bmat == 'I') { wnorm = igraphdnrm2_(n, &resid[1], &c__1); } /* %-----------------------------------------% */ /* | Compute the j-th residual corresponding | */ /* | to the j step factorization. | */ /* | Use Classical Gram Schmidt and compute: | */ /* | w_{j} <- V_{j}^T * B * OP * v_{j} | */ /* | r_{j} <- OP*v_{j} - V_{j} * w_{j} | */ /* %-----------------------------------------% */ /* %------------------------------------------% */ /* | Compute the j Fourier coefficients w_{j} | */ /* | WORKD(IPJ:IPJ+N-1) contains B*OP*v_{j}. | */ /* %------------------------------------------% */ igraphdgemv_("T", n, &j, &c_b25, &v[v_offset], ldv, &workd[ipj], &c__1, &c_b47, &h__[j * h_dim1 + 1], &c__1); /* %--------------------------------------% */ /* | Orthogonalize r_{j} against V_{j}. | */ /* | RESID contains OP*v_{j}. See STEP 3. | */ /* %--------------------------------------% */ igraphdgemv_("N", n, &j, &c_b50, &v[v_offset], ldv, &h__[j * h_dim1 + 1], &c__1, &c_b25, &resid[1], &c__1); if (j > 1) { h__[j + (j - 1) * h_dim1] = betaj; } igraphsecond_(&t4); orth1 = TRUE_; igraphsecond_(&t2); if (*(unsigned char *)bmat == 'G') { ++timing_1.nbx; igraphdcopy_(n, &resid[1], &c__1, &workd[irj], &c__1); ipntr[1] = irj; ipntr[2] = ipj; *ido = 2; /* %----------------------------------% */ /* | Exit in order to compute B*r_{j} | */ /* %----------------------------------% */ goto L9000; } else if (*(unsigned char *)bmat == 'I') { igraphdcopy_(n, &resid[1], &c__1, &workd[ipj], &c__1); } L70: /* %---------------------------------------------------% */ /* | Back from reverse communication if ORTH1 = .true. | */ /* | WORKD(IPJ:IPJ+N-1) := B*r_{j}. | */ /* %---------------------------------------------------% */ if (*(unsigned char *)bmat == 'G') { igraphsecond_(&t3); timing_1.tmvbx += t3 - t2; } orth1 = FALSE_; /* %------------------------------% */ /* | Compute the B-norm of r_{j}. | */ /* %------------------------------% */ if (*(unsigned char *)bmat == 'G') { *rnorm = igraphddot_(n, &resid[1], &c__1, &workd[ipj], &c__1); *rnorm = sqrt((abs(*rnorm))); } else if (*(unsigned char *)bmat == 'I') { *rnorm = igraphdnrm2_(n, &resid[1], &c__1); } /* %-----------------------------------------------------------% */ /* | STEP 5: Re-orthogonalization / Iterative refinement phase | */ /* | Maximum NITER_ITREF tries. | */ /* | | */ /* | s = V_{j}^T * B * r_{j} | */ /* | r_{j} = r_{j} - V_{j}*s | */ /* | alphaj = alphaj + s_{j} | */ /* | | */ /* | The stopping criteria used for iterative refinement is | */ /* | discussed in Parlett's book SEP, page 107 and in Gragg & | */ /* | Reichel ACM TOMS paper; Algorithm 686, Dec. 1990. | */ /* | Determine if we need to correct the residual. The goal is | */ /* | to enforce ||v(:,1:j)^T * r_{j}|| .le. eps * || r_{j} || | */ /* | The following test determines whether the sine of the | */ /* | angle between OP*x and the computed residual is less | */ /* | than or equal to 0.717. | */ /* %-----------------------------------------------------------% */ if (*rnorm > wnorm * .717f) { goto L100; } iter = 0; ++timing_1.nrorth; /* %---------------------------------------------------% */ /* | Enter the Iterative refinement phase. If further | */ /* | refinement is necessary, loop back here. The loop | */ /* | variable is ITER. Perform a step of Classical | */ /* | Gram-Schmidt using all the Arnoldi vectors V_{j} | */ /* %---------------------------------------------------% */ L80: if (msglvl > 2) { xtemp[0] = wnorm; xtemp[1] = *rnorm; igraphdvout_(&debug_1.logfil, &c__2, xtemp, &debug_1.ndigit, "_naitr: re-o" "rthonalization; wnorm and rnorm are"); igraphdvout_(&debug_1.logfil, &j, &h__[j * h_dim1 + 1], &debug_1.ndigit, "_naitr: j-th column of H"); } /* %----------------------------------------------------% */ /* | Compute V_{j}^T * B * r_{j}. | */ /* | WORKD(IRJ:IRJ+J-1) = v(:,1:J)'*WORKD(IPJ:IPJ+N-1). | */ /* %----------------------------------------------------% */ igraphdgemv_("T", n, &j, &c_b25, &v[v_offset], ldv, &workd[ipj], &c__1, &c_b47, &workd[irj], &c__1); /* %---------------------------------------------% */ /* | Compute the correction to the residual: | */ /* | r_{j} = r_{j} - V_{j} * WORKD(IRJ:IRJ+J-1). | */ /* | The correction to H is v(:,1:J)*H(1:J,1:J) | */ /* | + v(:,1:J)*WORKD(IRJ:IRJ+J-1)*e'_j. | */ /* %---------------------------------------------% */ igraphdgemv_("N", n, &j, &c_b50, &v[v_offset], ldv, &workd[irj], &c__1, &c_b25, &resid[1], &c__1); igraphdaxpy_(&j, &c_b25, &workd[irj], &c__1, &h__[j * h_dim1 + 1], &c__1); orth2 = TRUE_; igraphsecond_(&t2); if (*(unsigned char *)bmat == 'G') { ++timing_1.nbx; igraphdcopy_(n, &resid[1], &c__1, &workd[irj], &c__1); ipntr[1] = irj; ipntr[2] = ipj; *ido = 2; /* %-----------------------------------% */ /* | Exit in order to compute B*r_{j}. | */ /* | r_{j} is the corrected residual. | */ /* %-----------------------------------% */ goto L9000; } else if (*(unsigned char *)bmat == 'I') { igraphdcopy_(n, &resid[1], &c__1, &workd[ipj], &c__1); } L90: /* %---------------------------------------------------% */ /* | Back from reverse communication if ORTH2 = .true. | */ /* %---------------------------------------------------% */ if (*(unsigned char *)bmat == 'G') { igraphsecond_(&t3); timing_1.tmvbx += t3 - t2; } /* %-----------------------------------------------------% */ /* | Compute the B-norm of the corrected residual r_{j}. | */ /* %-----------------------------------------------------% */ if (*(unsigned char *)bmat == 'G') { rnorm1 = igraphddot_(n, &resid[1], &c__1, &workd[ipj], &c__1); rnorm1 = sqrt((abs(rnorm1))); } else if (*(unsigned char *)bmat == 'I') { rnorm1 = igraphdnrm2_(n, &resid[1], &c__1); } if (msglvl > 0 && iter > 0) { igraphivout_(&debug_1.logfil, &c__1, &j, &debug_1.ndigit, "_naitr: Iterati" "ve refinement for Arnoldi residual"); if (msglvl > 2) { xtemp[0] = *rnorm; xtemp[1] = rnorm1; igraphdvout_(&debug_1.logfil, &c__2, xtemp, &debug_1.ndigit, "_naitr: " "iterative refinement ; rnorm and rnorm1 are"); } } /* %-----------------------------------------% */ /* | Determine if we need to perform another | */ /* | step of re-orthogonalization. | */ /* %-----------------------------------------% */ if (rnorm1 > *rnorm * .717f) { /* %---------------------------------------% */ /* | No need for further refinement. | */ /* | The cosine of the angle between the | */ /* | corrected residual vector and the old | */ /* | residual vector is greater than 0.717 | */ /* | In other words the corrected residual | */ /* | and the old residual vector share an | */ /* | angle of less than arcCOS(0.717) | */ /* %---------------------------------------% */ *rnorm = rnorm1; } else { /* %-------------------------------------------% */ /* | Another step of iterative refinement step | */ /* | is required. NITREF is used by stat.h | */ /* %-------------------------------------------% */ ++timing_1.nitref; *rnorm = rnorm1; ++iter; if (iter <= 1) { goto L80; } /* %-------------------------------------------------% */ /* | Otherwise RESID is numerically in the span of V | */ /* %-------------------------------------------------% */ i__1 = *n; for (jj = 1; jj <= i__1; ++jj) { resid[jj] = 0.; /* L95: */ } *rnorm = 0.; } /* %----------------------------------------------% */ /* | Branch here directly if iterative refinement | */ /* | wasn't necessary or after at most NITER_REF | */ /* | steps of iterative refinement. | */ /* %----------------------------------------------% */ L100: rstart = FALSE_; orth2 = FALSE_; igraphsecond_(&t5); timing_1.titref += t5 - t4; /* %------------------------------------% */ /* | STEP 6: Update j = j+1; Continue | */ /* %------------------------------------% */ ++j; if (j > *k + *np) { igraphsecond_(&t1); timing_1.tnaitr += t1 - t0; *ido = 99; i__1 = *k + *np - 1; for (i__ = max(1,*k); i__ <= i__1; ++i__) { /* %--------------------------------------------% */ /* | Check for splitting and deflation. | */ /* | Use a standard test as in the QR algorithm | */ /* | REFERENCE: LAPACK subroutine dlahqr | */ /* %--------------------------------------------% */ tst1 = (d__1 = h__[i__ + i__ * h_dim1], abs(d__1)) + (d__2 = h__[ i__ + 1 + (i__ + 1) * h_dim1], abs(d__2)); if (tst1 == 0.) { i__2 = *k + *np; tst1 = igraphdlanhs_("1", &i__2, &h__[h_offset], ldh, &workd[*n + 1]); } /* Computing MAX */ d__2 = ulp * tst1; if ((d__1 = h__[i__ + 1 + i__ * h_dim1], abs(d__1)) <= max(d__2, smlnum)) { h__[i__ + 1 + i__ * h_dim1] = 0.; } /* L110: */ } if (msglvl > 2) { i__1 = *k + *np; i__2 = *k + *np; igraphdmout_(&debug_1.logfil, &i__1, &i__2, &h__[h_offset], ldh, & debug_1.ndigit, "_naitr: Final upper Hessenberg matrix H" " of order K+NP"); } goto L9000; } /* %--------------------------------------------------------% */ /* | Loop back to extend the factorization by another step. | */ /* %--------------------------------------------------------% */ goto L1000; /* %---------------------------------------------------------------% */ /* | | */ /* | E N D O F M A I N I T E R A T I O N L O O P | */ /* | | */ /* %---------------------------------------------------------------% */ L9000: return 0; /* %---------------% */ /* | End of igraphdnaitr | */ /* %---------------% */ } /* igraphdnaitr_ */
Subroutine */ int igraphdsaup2_(integer *ido, char *bmat, integer *n, char * which, integer *nev, integer *np, doublereal *tol, doublereal *resid, integer *mode, integer *iupd, integer *ishift, integer *mxiter, doublereal *v, integer *ldv, doublereal *h__, integer *ldh, doublereal *ritz, doublereal *bounds, doublereal *q, integer *ldq, doublereal *workl, integer *ipntr, doublereal *workd, integer *info) { /* System generated locals */ integer h_dim1, h_offset, q_dim1, q_offset, v_dim1, v_offset, i__1, i__2, i__3; doublereal d__1, d__2, d__3; /* Builtin functions */ double pow_dd(doublereal *, doublereal *); integer s_cmp(char *, char *, ftnlen, ftnlen); /* Subroutine */ int s_copy(char *, char *, ftnlen, ftnlen); double sqrt(doublereal); /* Local variables */ integer j; real t0, t1, t2, t3; integer kp[3]; IGRAPH_F77_SAVE integer np0; integer nbx = 0; IGRAPH_F77_SAVE integer nev0; extern doublereal igraphddot_(integer *, doublereal *, integer *, doublereal *, integer *); IGRAPH_F77_SAVE doublereal eps23; integer ierr; IGRAPH_F77_SAVE integer iter; doublereal temp; integer nevd2; extern doublereal igraphdnrm2_(integer *, doublereal *, integer *); IGRAPH_F77_SAVE logical getv0; integer nevm2; IGRAPH_F77_SAVE logical cnorm; extern /* Subroutine */ int igraphdcopy_(integer *, doublereal *, integer *, doublereal *, integer *), igraphdswap_(integer *, doublereal *, integer *, doublereal *, integer *); IGRAPH_F77_SAVE integer nconv; IGRAPH_F77_SAVE logical initv; IGRAPH_F77_SAVE doublereal rnorm; real tmvbx = 0.0; extern /* Subroutine */ int igraphdvout_(integer *, integer *, doublereal *, integer *, char *, ftnlen), igraphivout_(integer *, integer *, integer * , integer *, char *, ftnlen), igraphdgetv0_(integer *, char *, integer * , logical *, integer *, integer *, doublereal *, integer *, doublereal *, doublereal *, integer *, doublereal *, integer *); integer msaup2 = 0; real tsaup2; extern doublereal igraphdlamch_(char *); integer nevbef; extern /* Subroutine */ int igraphsecond_(real *); integer logfil, ndigit; extern /* Subroutine */ int igraphdseigt_(doublereal *, integer *, doublereal *, integer *, doublereal *, doublereal *, doublereal *, integer *); IGRAPH_F77_SAVE logical update; extern /* Subroutine */ int igraphdsaitr_(integer *, char *, integer *, integer *, integer *, integer *, doublereal *, doublereal *, doublereal *, integer *, doublereal *, integer *, integer *, doublereal *, integer *), igraphdsgets_(integer *, char *, integer *, integer *, doublereal *, doublereal *, doublereal *), igraphdsapps_( integer *, integer *, integer *, doublereal *, doublereal *, integer *, doublereal *, integer *, doublereal *, doublereal *, integer *, doublereal *), igraphdsconv_(integer *, doublereal *, doublereal *, doublereal *, integer *); IGRAPH_F77_SAVE logical ushift; char wprime[2]; IGRAPH_F77_SAVE integer msglvl; integer nptemp; extern /* Subroutine */ int igraphdsortr_(char *, logical *, integer *, doublereal *, doublereal *); IGRAPH_F77_SAVE integer kplusp; /* %----------------------------------------------------% | Include files for debugging and timing information | %----------------------------------------------------% %------------------% | Scalar Arguments | %------------------% %-----------------% | Array Arguments | %-----------------% %------------% | Parameters | %------------% %---------------% | Local Scalars | %---------------% %----------------------% | External Subroutines | %----------------------% %--------------------% | External Functions | %--------------------% %---------------------% | Intrinsic Functions | %---------------------% %-----------------------% | Executable Statements | %-----------------------% Parameter adjustments */ --workd; --resid; --workl; --bounds; --ritz; v_dim1 = *ldv; v_offset = 1 + v_dim1; v -= v_offset; h_dim1 = *ldh; h_offset = 1 + h_dim1; h__ -= h_offset; q_dim1 = *ldq; q_offset = 1 + q_dim1; q -= q_offset; --ipntr; /* Function Body */ if (*ido == 0) { /* %-------------------------------% | Initialize timing statistics | | & message level for debugging | %-------------------------------% */ igraphsecond_(&t0); msglvl = msaup2; /* %---------------------------------% | Set machine dependent constant. | %---------------------------------% */ eps23 = igraphdlamch_("Epsilon-Machine"); eps23 = pow_dd(&eps23, &c_b3); /* %-------------------------------------% | nev0 and np0 are integer variables | | hold the initial values of NEV & NP | %-------------------------------------% */ nev0 = *nev; np0 = *np; /* %-------------------------------------% | kplusp is the bound on the largest | | Lanczos factorization built. | | nconv is the current number of | | "converged" eigenvlues. | | iter is the counter on the current | | iteration step. | %-------------------------------------% */ kplusp = nev0 + np0; nconv = 0; iter = 0; /* %--------------------------------------------% | Set flags for computing the first NEV steps | | of the Lanczos factorization. | %--------------------------------------------% */ getv0 = TRUE_; update = FALSE_; ushift = FALSE_; cnorm = FALSE_; if (*info != 0) { /* %--------------------------------------------% | User provides the initial residual vector. | %--------------------------------------------% */ initv = TRUE_; *info = 0; } else { initv = FALSE_; } } /* %---------------------------------------------% | Get a possibly random starting vector and | | force it into the range of the operator OP. | %---------------------------------------------% L10: */ if (getv0) { igraphdgetv0_(ido, bmat, &c__1, &initv, n, &c__1, &v[v_offset], ldv, &resid[ 1], &rnorm, &ipntr[1], &workd[1], info); if (*ido != 99) { goto L9000; } if (rnorm == 0.) { /* %-----------------------------------------% | The initial vector is zero. Error exit. | %-----------------------------------------% */ *info = -9; goto L1200; } getv0 = FALSE_; *ido = 0; } /* %------------------------------------------------------------% | Back from reverse communication: continue with update step | %------------------------------------------------------------% */ if (update) { goto L20; } /* %-------------------------------------------% | Back from computing user specified shifts | %-------------------------------------------% */ if (ushift) { goto L50; } /* %-------------------------------------% | Back from computing residual norm | | at the end of the current iteration | %-------------------------------------% */ if (cnorm) { goto L100; } /* %----------------------------------------------------------% | Compute the first NEV steps of the Lanczos factorization | %----------------------------------------------------------% */ igraphdsaitr_(ido, bmat, n, &c__0, &nev0, mode, &resid[1], &rnorm, &v[v_offset], ldv, &h__[h_offset], ldh, &ipntr[1], &workd[1], info); /* %---------------------------------------------------% | ido .ne. 99 implies use of reverse communication | | to compute operations involving OP and possibly B | %---------------------------------------------------% */ if (*ido != 99) { goto L9000; } if (*info > 0) { /* %-----------------------------------------------------% | dsaitr was unable to build an Lanczos factorization | | of length NEV0. INFO is returned with the size of | | the factorization built. Exit main loop. | %-----------------------------------------------------% */ *np = *info; *mxiter = iter; *info = -9999; goto L1200; } /* %--------------------------------------------------------------% | | | M A I N LANCZOS I T E R A T I O N L O O P | | Each iteration implicitly restarts the Lanczos | | factorization in place. | | | %--------------------------------------------------------------% */ L1000: ++iter; if (msglvl > 0) { igraphivout_(&logfil, &c__1, &iter, &ndigit, "_saup2: **** Start of major " "iteration number ****", (ftnlen)49); } if (msglvl > 1) { igraphivout_(&logfil, &c__1, nev, &ndigit, "_saup2: The length of the curr" "ent Lanczos factorization", (ftnlen)55); igraphivout_(&logfil, &c__1, np, &ndigit, "_saup2: Extend the Lanczos fact" "orization by", (ftnlen)43); } /* %------------------------------------------------------------% | Compute NP additional steps of the Lanczos factorization. | %------------------------------------------------------------% */ *ido = 0; L20: update = TRUE_; igraphdsaitr_(ido, bmat, n, nev, np, mode, &resid[1], &rnorm, &v[v_offset], ldv, &h__[h_offset], ldh, &ipntr[1], &workd[1], info); /* %---------------------------------------------------% | ido .ne. 99 implies use of reverse communication | | to compute operations involving OP and possibly B | %---------------------------------------------------% */ if (*ido != 99) { goto L9000; } if (*info > 0) { /* %-----------------------------------------------------% | dsaitr was unable to build an Lanczos factorization | | of length NEV0+NP0. INFO is returned with the size | | of the factorization built. Exit main loop. | %-----------------------------------------------------% */ *np = *info; *mxiter = iter; *info = -9999; goto L1200; } update = FALSE_; if (msglvl > 1) { igraphdvout_(&logfil, &c__1, &rnorm, &ndigit, "_saup2: Current B-norm of r" "esidual for factorization", (ftnlen)52); } /* %--------------------------------------------------------% | Compute the eigenvalues and corresponding error bounds | | of the current symmetric tridiagonal matrix. | %--------------------------------------------------------% */ igraphdseigt_(&rnorm, &kplusp, &h__[h_offset], ldh, &ritz[1], &bounds[1], & workl[1], &ierr); if (ierr != 0) { *info = -8; goto L1200; } /* %----------------------------------------------------% | Make a copy of eigenvalues and corresponding error | | bounds obtained from _seigt. | %----------------------------------------------------% */ igraphdcopy_(&kplusp, &ritz[1], &c__1, &workl[kplusp + 1], &c__1); igraphdcopy_(&kplusp, &bounds[1], &c__1, &workl[(kplusp << 1) + 1], &c__1); /* %---------------------------------------------------% | Select the wanted Ritz values and their bounds | | to be used in the convergence test. | | The selection is based on the requested number of | | eigenvalues instead of the current NEV and NP to | | prevent possible misconvergence. | | * Wanted Ritz values := RITZ(NP+1:NEV+NP) | | * Shifts := RITZ(1:NP) := WORKL(1:NP) | %---------------------------------------------------% */ *nev = nev0; *np = np0; igraphdsgets_(ishift, which, nev, np, &ritz[1], &bounds[1], &workl[1]); /* %-------------------% | Convergence test. | %-------------------% */ igraphdcopy_(nev, &bounds[*np + 1], &c__1, &workl[*np + 1], &c__1); igraphdsconv_(nev, &ritz[*np + 1], &workl[*np + 1], tol, &nconv); if (msglvl > 2) { kp[0] = *nev; kp[1] = *np; kp[2] = nconv; igraphivout_(&logfil, &c__3, kp, &ndigit, "_saup2: NEV, NP, NCONV are", ( ftnlen)26); igraphdvout_(&logfil, &kplusp, &ritz[1], &ndigit, "_saup2: The eigenvalues" " of H", (ftnlen)28); igraphdvout_(&logfil, &kplusp, &bounds[1], &ndigit, "_saup2: Ritz estimate" "s of the current NCV Ritz values", (ftnlen)53); } /* %---------------------------------------------------------% | Count the number of unwanted Ritz values that have zero | | Ritz estimates. If any Ritz estimates are equal to zero | | then a leading block of H of order equal to at least | | the number of Ritz values with zero Ritz estimates has | | split off. None of these Ritz values may be removed by | | shifting. Decrease NP the number of shifts to apply. If | | no shifts may be applied, then prepare to exit | %---------------------------------------------------------% */ nptemp = *np; i__1 = nptemp; for (j = 1; j <= i__1; ++j) { if (bounds[j] == 0.) { --(*np); ++(*nev); } /* L30: */ } if (nconv >= nev0 || iter > *mxiter || *np == 0) { /* %------------------------------------------------% | Prepare to exit. Put the converged Ritz values | | and corresponding bounds in RITZ(1:NCONV) and | | BOUNDS(1:NCONV) respectively. Then sort. Be | | careful when NCONV > NP since we don't want to | | swap overlapping locations. | %------------------------------------------------% */ if (s_cmp(which, "BE", (ftnlen)2, (ftnlen)2) == 0) { /* %-----------------------------------------------------% | Both ends of the spectrum are requested. | | Sort the eigenvalues into algebraically decreasing | | order first then swap low end of the spectrum next | | to high end in appropriate locations. | | NOTE: when np < floor(nev/2) be careful not to swap | | overlapping locations. | %-----------------------------------------------------% */ s_copy(wprime, "SA", (ftnlen)2, (ftnlen)2); igraphdsortr_(wprime, &c_true, &kplusp, &ritz[1], &bounds[1]) ; nevd2 = *nev / 2; nevm2 = *nev - nevd2; if (*nev > 1) { i__1 = min(nevd2,*np); /* Computing MAX */ i__2 = kplusp - nevd2 + 1, i__3 = kplusp - *np + 1; igraphdswap_(&i__1, &ritz[nevm2 + 1], &c__1, &ritz[max(i__2,i__3)], &c__1); i__1 = min(nevd2,*np); /* Computing MAX */ i__2 = kplusp - nevd2 + 1, i__3 = kplusp - *np; igraphdswap_(&i__1, &bounds[nevm2 + 1], &c__1, &bounds[max(i__2, i__3) + 1], &c__1); } } else { /* %--------------------------------------------------% | LM, SM, LA, SA case. | | Sort the eigenvalues of H into the an order that | | is opposite to WHICH, and apply the resulting | | order to BOUNDS. The eigenvalues are sorted so | | that the wanted part are always within the first | | NEV locations. | %--------------------------------------------------% */ if (s_cmp(which, "LM", (ftnlen)2, (ftnlen)2) == 0) { s_copy(wprime, "SM", (ftnlen)2, (ftnlen)2); } if (s_cmp(which, "SM", (ftnlen)2, (ftnlen)2) == 0) { s_copy(wprime, "LM", (ftnlen)2, (ftnlen)2); } if (s_cmp(which, "LA", (ftnlen)2, (ftnlen)2) == 0) { s_copy(wprime, "SA", (ftnlen)2, (ftnlen)2); } if (s_cmp(which, "SA", (ftnlen)2, (ftnlen)2) == 0) { s_copy(wprime, "LA", (ftnlen)2, (ftnlen)2); } igraphdsortr_(wprime, &c_true, &kplusp, &ritz[1], &bounds[1]) ; } /* %--------------------------------------------------% | Scale the Ritz estimate of each Ritz value | | by 1 / max(eps23,magnitude of the Ritz value). | %--------------------------------------------------% */ i__1 = nev0; for (j = 1; j <= i__1; ++j) { /* Computing MAX */ d__2 = eps23, d__3 = (d__1 = ritz[j], abs(d__1)); temp = max(d__2,d__3); bounds[j] /= temp; /* L35: */ } /* %----------------------------------------------------% | Sort the Ritz values according to the scaled Ritz | | esitmates. This will push all the converged ones | | towards the front of ritzr, ritzi, bounds | | (in the case when NCONV < NEV.) | %----------------------------------------------------% */ s_copy(wprime, "LA", (ftnlen)2, (ftnlen)2); igraphdsortr_(wprime, &c_true, &nev0, &bounds[1], &ritz[1]); /* %----------------------------------------------% | Scale the Ritz estimate back to its original | | value. | %----------------------------------------------% */ i__1 = nev0; for (j = 1; j <= i__1; ++j) { /* Computing MAX */ d__2 = eps23, d__3 = (d__1 = ritz[j], abs(d__1)); temp = max(d__2,d__3); bounds[j] *= temp; /* L40: */ } /* %--------------------------------------------------% | Sort the "converged" Ritz values again so that | | the "threshold" values and their associated Ritz | | estimates appear at the appropriate position in | | ritz and bound. | %--------------------------------------------------% */ if (s_cmp(which, "BE", (ftnlen)2, (ftnlen)2) == 0) { /* %------------------------------------------------% | Sort the "converged" Ritz values in increasing | | order. The "threshold" values are in the | | middle. | %------------------------------------------------% */ s_copy(wprime, "LA", (ftnlen)2, (ftnlen)2); igraphdsortr_(wprime, &c_true, &nconv, &ritz[1], &bounds[1]); } else { /* %----------------------------------------------% | In LM, SM, LA, SA case, sort the "converged" | | Ritz values according to WHICH so that the | | "threshold" value appears at the front of | | ritz. | %----------------------------------------------% */ igraphdsortr_(which, &c_true, &nconv, &ritz[1], &bounds[1]); } /* %------------------------------------------% | Use h( 1,1 ) as storage to communicate | | rnorm to _seupd if needed | %------------------------------------------% */ h__[h_dim1 + 1] = rnorm; if (msglvl > 1) { igraphdvout_(&logfil, &kplusp, &ritz[1], &ndigit, "_saup2: Sorted Ritz" " values.", (ftnlen)27); igraphdvout_(&logfil, &kplusp, &bounds[1], &ndigit, "_saup2: Sorted ri" "tz estimates.", (ftnlen)30); } /* %------------------------------------% | Max iterations have been exceeded. | %------------------------------------% */ if (iter > *mxiter && nconv < *nev) { *info = 1; } /* %---------------------% | No shifts to apply. | %---------------------% */ if (*np == 0 && nconv < nev0) { *info = 2; } *np = nconv; goto L1100; } else if (nconv < *nev && *ishift == 1) { /* %---------------------------------------------------% | Do not have all the requested eigenvalues yet. | | To prevent possible stagnation, adjust the number | | of Ritz values and the shifts. | %---------------------------------------------------% */ nevbef = *nev; /* Computing MIN */ i__1 = nconv, i__2 = *np / 2; *nev += min(i__1,i__2); if (*nev == 1 && kplusp >= 6) { *nev = kplusp / 2; } else if (*nev == 1 && kplusp > 2) { *nev = 2; } *np = kplusp - *nev; /* %---------------------------------------% | If the size of NEV was just increased | | resort the eigenvalues. | %---------------------------------------% */ if (nevbef < *nev) { igraphdsgets_(ishift, which, nev, np, &ritz[1], &bounds[1], &workl[1]); } } if (msglvl > 0) { igraphivout_(&logfil, &c__1, &nconv, &ndigit, "_saup2: no. of \"converge" "d\" Ritz values at this iter.", (ftnlen)52); if (msglvl > 1) { kp[0] = *nev; kp[1] = *np; igraphivout_(&logfil, &c__2, kp, &ndigit, "_saup2: NEV and NP are", ( ftnlen)22); igraphdvout_(&logfil, nev, &ritz[*np + 1], &ndigit, "_saup2: \"wante" "d\" Ritz values.", (ftnlen)29); igraphdvout_(&logfil, nev, &bounds[*np + 1], &ndigit, "_saup2: Ritz es" "timates of the \"wanted\" values ", (ftnlen)46); } } if (*ishift == 0) { /* %-----------------------------------------------------% | User specified shifts: reverse communication to | | compute the shifts. They are returned in the first | | NP locations of WORKL. | %-----------------------------------------------------% */ ushift = TRUE_; *ido = 3; goto L9000; } L50: /* %------------------------------------% | Back from reverse communication; | | User specified shifts are returned | | in WORKL(1:*NP) | %------------------------------------% */ ushift = FALSE_; /* %---------------------------------------------------------% | Move the NP shifts to the first NP locations of RITZ to | | free up WORKL. This is for the non-exact shift case; | | in the exact shift case, dsgets already handles this. | %---------------------------------------------------------% */ if (*ishift == 0) { igraphdcopy_(np, &workl[1], &c__1, &ritz[1], &c__1); } if (msglvl > 2) { igraphivout_(&logfil, &c__1, np, &ndigit, "_saup2: The number of shifts to" " apply ", (ftnlen)38); igraphdvout_(&logfil, np, &workl[1], &ndigit, "_saup2: shifts selected", ( ftnlen)23); if (*ishift == 1) { igraphdvout_(&logfil, np, &bounds[1], &ndigit, "_saup2: corresponding " "Ritz estimates", (ftnlen)36); } } /* %---------------------------------------------------------% | Apply the NP0 implicit shifts by QR bulge chasing. | | Each shift is applied to the entire tridiagonal matrix. | | The first 2*N locations of WORKD are used as workspace. | | After dsapps is done, we have a Lanczos | | factorization of length NEV. | %---------------------------------------------------------% */ igraphdsapps_(n, nev, np, &ritz[1], &v[v_offset], ldv, &h__[h_offset], ldh, & resid[1], &q[q_offset], ldq, &workd[1]); /* %---------------------------------------------% | Compute the B-norm of the updated residual. | | Keep B*RESID in WORKD(1:N) to be used in | | the first step of the next call to dsaitr. | %---------------------------------------------% */ cnorm = TRUE_; igraphsecond_(&t2); if (*(unsigned char *)bmat == 'G') { ++nbx; igraphdcopy_(n, &resid[1], &c__1, &workd[*n + 1], &c__1); ipntr[1] = *n + 1; ipntr[2] = 1; *ido = 2; /* %----------------------------------% | Exit in order to compute B*RESID | %----------------------------------% */ goto L9000; } else if (*(unsigned char *)bmat == 'I') { igraphdcopy_(n, &resid[1], &c__1, &workd[1], &c__1); } L100: /* %----------------------------------% | Back from reverse communication; | | WORKD(1:N) := B*RESID | %----------------------------------% */ if (*(unsigned char *)bmat == 'G') { igraphsecond_(&t3); tmvbx += t3 - t2; } if (*(unsigned char *)bmat == 'G') { rnorm = igraphddot_(n, &resid[1], &c__1, &workd[1], &c__1); rnorm = sqrt((abs(rnorm))); } else if (*(unsigned char *)bmat == 'I') { rnorm = igraphdnrm2_(n, &resid[1], &c__1); } cnorm = FALSE_; /* L130: */ if (msglvl > 2) { igraphdvout_(&logfil, &c__1, &rnorm, &ndigit, "_saup2: B-norm of residual " "for NEV factorization", (ftnlen)48); igraphdvout_(&logfil, nev, &h__[(h_dim1 << 1) + 1], &ndigit, "_saup2: main" " diagonal of compressed H matrix", (ftnlen)44); i__1 = *nev - 1; igraphdvout_(&logfil, &i__1, &h__[h_dim1 + 2], &ndigit, "_saup2: subdiagon" "al of compressed H matrix", (ftnlen)42); } goto L1000; /* %---------------------------------------------------------------% | | | E N D O F M A I N I T E R A T I O N L O O P | | | %---------------------------------------------------------------% */ L1100: *mxiter = iter; *nev = nconv; L1200: *ido = 99; /* %------------% | Error exit | %------------% */ igraphsecond_(&t1); tsaup2 = t1 - t0; L9000: return 0; /* %---------------% | End of dsaup2 | %---------------% */ } /* igraphdsaup2_ */